A High-Resolution Landscape Dataset for Concept-Based XAI With Application to Species Distribution Models

1. Problem Statement

Species Distribution Models (SDMs) are critical for conservation policy and invasive species management. They traditionally serve a dual purpose: achieving high predictive accuracy and providing ecological insights into the drivers of species distribution.

• The Challenge: While deep learning (DL) models have significantly improved predictive performance over traditional machine learning (e.g., Random Forests) by leveraging high-dimensional remote sensing data, they act as "black boxes." Their abstract feature learning across millions of parameters makes it difficult to extract interpretable ecological knowledge.
• The Gap: Existing Explainable AI (XAI) methods (like LIME or SHAP) often operate at the pixel level, which is difficult for ecologists to interpret. Furthermore, concept-based XAI (which explains models using high-level human concepts like "hedgerows" or "wetlands") has not been implemented for SDMs because a dedicated, high-resolution landscape concept dataset did not previously exist.

2. Methodology

The authors propose a framework to reconcile predictive performance with interpretability using Robust TCAV (Testing with Concept Activation Vectors).

A. Dataset Creation

The core contribution is the creation of a new, open-access dataset derived from high-resolution drone imagery (multispectral and LiDAR).

• Data Source: Acquired in April 2024 across five study sites in France using a Trinity F90+ drone equipped with a MicaSense Dual MX sensor (5-band multispectral) and a Qube240 LiDAR sensor.
• Resolution: 8 cm/pixel (multispectral) and 10 cm/pixel (LiDAR-derived elevation models).
• Concept Patches: 653 patches representing 15 distinct landscape concepts (e.g., Hedgerows, Isolated Trees, Wetlands, Roads, various crops). These were manually mapped using hydrographic networks, agricultural registers, and photointerpretation.
• Reference Patches: 1,450 random patches to serve as negative controls.
• Species Data: Distribution data for two aquatic insect taxa (Plecoptera and Trichoptera) collected via 234 sticky traps. Labels are binary: Presence vs. Absence.

B. Model Architectures

Three distinct deep learning models were trained to predict insect presence/absence using 7-channel inputs (5 multispectral bands + 2 LiDAR elevation models):

1. CerberusCNN: A custom CNN with three parallel branches using different kernel sizes (3, 5, 7) to capture multi-scale spatial features. (~270k parameters).
2. Adapted ResNet-50 (ARN-50): A standard deep CNN benchmark, modified to accept 7-channel inputs. (~23.5M parameters).
3. PicoViT: A lightweight Vision Transformer adapted for the dataset scale, using 3 layers and 3 attention heads. (~146k parameters).

C. Interpretation Framework: Robust TCAV

Instead of analyzing pixel importance, the authors applied Robust TCAV to quantify the influence of the 15 landscape concepts on model predictions.

• Mechanism: The method computes a Concept Activation Vector (CAV) by finding the direction in the model's latent space that separates concept patches from random patches.
• Scoring: It calculates the sensitivity of the model's prediction to the concept. A score > 0.5 indicates the concept positively influences the prediction; < 0.5 indicates a negative influence.
• Robustness: The authors used the "Robust" variant (using mean activation differences rather than hyperplane normals) and tested different thresholds (finding a zero-threshold more suitable for their ecological data).

3. Key Contributions

1. First Concept-Based XAI Dataset for SDMs: The release of a high-resolution, multi-modal (spectral + LiDAR) dataset containing 15 ecologically relevant landscape concepts and 1,450 random references.
2. Novel SDMs: Development and training of three distinct architectures (CerberusCNN, ARN-50, PicoViT) specifically adapted for high-resolution drone data to predict aquatic insect distributions.
3. First Implementation of Robust TCAV in SDMs: Demonstrating that concept-based XAI can validate model logic against expert ecological knowledge and uncover novel associations.
4. Open Resources: Public release of the concept dataset, species distribution data, and code.

4. Results

Predictive Performance

• Best Performer: CerberusCNN achieved the most reliable results with the lowest standard deviation across taxa (Plecoptera AUC: 0.88; Trichoptera AUC: 0.79).
• Pretrained Models: Pretrained ARN-50 performed exceptionally well for Plecoptera (AUC: 0.90) but less consistently for Trichoptera.
• Transformers: PicoViT performed well for Plecoptera (AUC: 0.72) but failed to provide reliable predictions for Trichoptera (AUC: 0.66), suggesting that local pixel neighborhoods (captured by CNNs) may be more informative for these insects than long-range dependencies (captured by Transformers).

Interpretability (Robust TCAV)

• Ecological Alignment: The models generally aligned with expert knowledge.
  • Positive Drivers: Natural features (Woodlands, Wetlands, Isolated Trees, Permanent Grasslands) consistently showed high TCAV scores (>0.85) for Presence.
  • Negative Drivers: Anthropogenic features (Roads, Buildings, Crops) consistently showed low scores (<0.16) for Presence, indicating they drive Absence.
• Nuanced Findings:
  • Organic vs. Conventional: While both crop types negatively influenced presence, organic crops consistently ranked higher (less negative) than conventional crops, supporting the hypothesis that organic farming better sustains biodiversity.
  • Counter-Intuitive Results: For Trichoptera, models (CerberusCNN and ARN-50) surprisingly linked Roads and Buildings to Presence. This contradicts some literature but suggests a need for further ecological investigation, highlighting XAI's ability to generate new hypotheses.
  • Model Consistency: Most concepts showed concordant importance across architectures (12 out of 15 concepts). However, PicoViT produced near-zero scores for most concepts, reinforcing that unreliable models yield meaningless ecological insights.

5. Significance and Conclusion

• Bridging the Gap: The study successfully demonstrates that concept-based XAI can bridge the gap between the high predictive power of deep learning and the need for ecological interpretability.
• Policy Impact: By providing clear, class-level explanations (e.g., "Wetlands drive presence"), the framework empowers policymakers and land managers to make informed, landscape-scale decisions rather than relying on opaque black-box predictions.
• Validation: The results validate that deep learning SDMs are not just predictive tools but can be "opened up" to confirm known ecological drivers and discover new ones.
• Future Work: The authors suggest expanding to other concept-based XAI methods (e.g., Concept Activation Regions) and exploring why certain models (like PicoViT) fail to capture specific ecological signals in this context.

In summary, this paper provides a foundational dataset and methodology for making deep learning-based species distribution models transparent, trustworthy, and actionable for ecological conservation.